When something looks wrong

Parseh says what went wrong in words, on the page where it happened: in a note under a greyed-out button, in a message at the foot of the window, in a badge beside a build, in a dialog. The tables below quote those words as the pages show them, so you can find yours with the guide’s search box or your browser’s find-in-page, and say what each one means and what to do next. A message in italics is what a page says, one in code what a terminal prints; a label in bold is a button or a link.

When a message is not here, read it twice: most of them name the file, the rule or the button they are about.

1. Starting Parseh and installing it#

What you see	What it means, and what to do
The browser says the connection is not private, or the certificate is not trusted.	Expected, once per browser. Parseh makes its own certificate the first time it starts, and nobody else vouches for it — it is your server on your network. Choose Advanced and go on to the address. If the machine’s addresses have changed (a new network, a new Tailscale address) and a device refuses the old certificate, make a fresh one: `./serve.sh cert` (Windows: `serve.bat cert`), then restart.
A page will not load at all.	The server is not running, or it runs on another port. On Linux and macOS, `./serve.sh status` says `running: pid …, https port …` or `not running`, and `./serve.sh log` shows its last lines. On Windows the server runs in its own window: that window is the log. Start it again with `./serve.sh`, Parseh.command on a Mac, or a double-click on serve.bat.
Parseh stopped. Start it again on the computer it runs on, the way you started it before (Parseh.command on a Mac, serve.bat on Windows), and reload this page. (the studio says Server stopped)	Somebody pressed a stop button: ⏻ stop on the hub, ⏻ in the video player, stop server in a book’s reader, Stop server in the studio and the exercises. Nothing is lost. Start Parseh again as above — on Windows with a double-click on serve.bat — and reload the page.
Stopping asks 1 task is still running (or N tasks are) and lists them.	The server is still doing something you started — a build, an upload, a backup, a download being packed. Stopping now cuts it off. Answer Cancel, wait for the Working… pill to go, then stop.
`./serve.sh` says `already running: pid … on port …`	A server is already up. Use it, or `./serve.sh restart` to replace it (after changing a file of Parseh’s by hand, say; an update from Settings restarts it by itself).
`./serve.sh` says `failed to start -- the log says:` and quotes the log.	Usually the port is taken by another program (`Address already in use`). Start on another port — `./serve.sh 9000`, `serve.bat 9000` — or stop the other program.
`note: no 'ilya-frank' environment -- serving with …`	Serving works with any Python 3, but rebuilding books, dividing Japanese and Chinese into words and opening a modern Anki export need the packages of the environment. Run the installer once: `./install.sh`, Parseh.command on a Mac, install.bat on Windows.
Screen capture (the frame button on a card) does nothing, or copying to the clipboard fails.	The page is not on a secure address. Browsers only allow both on `https://` pages and on `localhost`. Use one of the `https://` addresses Parseh printed when it started. An old `http://` bookmark on the same port is sent to the `https://` address by itself; a server started with `--http` — or on Windows with no `openssl` to make a certificate, which the setup wizard says — serves plain http, where both work on `localhost` only.
The installer ends with `N checks passed, M missing.`	M counts the `MISS` lines, and each says what is missing and what it is for. Several matter for one job only: `ffmpeg` snaps a narration’s timings to its silences and cuts a card’s recording on the server, with its waveform — without it the timings are not snapped and the clip is recorded in the browser instead, as a WAV; `pdftotext` only re-extracts a book’s source. Something optional that is simply not there — LuaLaTeX, a dictionary, a parallel corpus, a translation model, a word analyzer — is a `--` line, and is not counted as missing. What serving needs is in the first part of the report, `required to serve the reader`.
The installation did not finish – the messages above say why. (install.bat; Parseh.command and `./install.sh` say …the lines above say why)	A step failed on the way, and the lines above it say which and why — often it was only the network, during a download. Mend what they name and double-click install.bat (or Parseh.command) again.
Parseh needs Python 3, and this computer does not have it. (serve.bat)	Windows found no Python 3.8 or newer. The simplest answer is the one it gives first: close the window and double-click install.bat, which brings its own Python with everything else Parseh needs. Or answer Y to Install Python 3.12 now? and winget installs it; where there is no winget, the download page of python.org opens.
Python was installed but cannot be found from here yet: close this window and double-click serve.bat again.	Just that: the window was opened before Python was there, and cannot see it. Close it and double-click serve.bat again.
`python3 does not run here (…).` (`./serve.sh`, usually on a Mac)	The Mac’s `python3` is only a stub until Apple’s command line tools are installed, and Parseh has no environment of its own yet. Double-click Parseh.command: the first time, it installs Parseh’s environment, with a Python of its own.
`openssl is needed to make the certificate (apt install openssl), or run with --http to serve without TLS` (quoted by `./serve.sh` after `failed to start`)	Parseh makes its certificate with `openssl`, and this machine has none. Install it with the system’s package manager, then start again; meanwhile `python3 serve.py --http` serves plain http. serve.bat never stops here: without `openssl` it serves plain http by itself, and its wizard offers to install one (Install OpenSSL with winget now?).
The installer says `no micromamba for …: install Python 3, or conda and then ./install.sh --conda`	There is no ready-made micromamba for this kind of machine. Install Python 3 (or conda) yourself and run the installer again; with conda, `./install.sh --conda` makes the environment with it.
`no Noto Serif CJK JP — a Japanese page falls back to another face …` (or the same for Chinese)	You have Japanese or Chinese content, and the font its pages and PDFs are set in is not on the machine. Install the Noto CJK fonts with your system’s package manager (the line says how); the pages work meanwhile, a PDF cannot build.
`tex: 'ngerman' is in no language.dat, so its books break at ENGLISH points` (or another language)	The hyphenation patterns are installed but TeX cannot see them, so a book in that language would break its words at English points. Run the two commands the line gives, then `./install.sh --pdf` again.
Parseh refused this: it came from another site … or … it came from a page with no address of its own …	A page that is not one of Parseh’s own — a page on another site, open in the same browser, or a page opened from the disk — tried to change something in Parseh, and nothing was changed: only Parseh’s own pages may. Reading is never refused. If it happened on a page you meant to use, open that page from Parseh itself — the hub, or one of the addresses Parseh printed when it started — and not from a copy saved to the disk, then press the button again.

More in Installing Parseh and Starting and stopping.

2. Updating Parseh#

What you see	What it means, and what to do
Which version am I running?	The foot of the hub says it (on a phone, on the very last line), and so does ⚙ Settings, and the first line the server writes when it starts. Give it in any report of a fault.
This Parseh has no list of its own files (.parseh-release.json): it was installed before Parseh could update itself, or unpacked by hand.	An update will not guess which files are Parseh’s and which are yours. Move into a fresh install once, by hand — a few folders — and from then on it updates from Settings (Moving into a fresh install, once).
This Parseh is a copy of its source code (a git checkout), not an install made from a release.	Git keeps that folder’s files, so Settings leaves them alone: bring it up to date with git, or install a release and move your things into it, as the row above.
No release of Parseh has been published on GitHub yet.	There is nothing newer to fetch. A release’s zip you have can still be chosen under A zip of your own.
GitHub could not be reached: … or GitHub asks Parseh to wait before asking again …	The computer has no network, or GitHub is refusing questions from this address for a while (it answers some sixty an hour). Nothing was changed; Check now again later.
That zip is not a release of Parseh …, That zip is damaged, or was changed after it was built: …	The file chosen is not a release’s zip, or a file in it does not match the zip’s own list. Download the release again from the releases page — the `parseh-<version>.zip` under Assets, not the Source code archive.
The button stays grey, under What may not survive going back.	Going back to that version would leave some of your data in a shape it may read wrongly. Read the list; if you still want to go back, tick I understand, and want to go back anyway: a copy of the small files of your books, videos, decks and settings is kept first.
A phone shows the page with a lock, and no button to install.	Installing another version changes what Parseh runs, so it is done on the computer Parseh runs on. The phone may still Check now.
The page says an update was undone.	The update was cut off and could not be finished, or the new version did not start: Parseh put back every file it had replaced, from the copies it keeps in `.parseh-update/`, and started the version you had. The report says what happened; nothing of yours was touched.
The update to … did not happen. It stopped while …	Something on this computer stopped it, and everything was put back as it was. The sentence names the step and what went wrong — the disk is full, this computer did not let it change a file, the release’s zip could not be read — so mend that (make room, give Parseh’s folder back to your user, download the zip again) and press the button again. The line in small type beneath is the technical one: copy it into a report of the fault.
Yours once, and not put back: … the copy it kept is gone	A file of yours that an update replaced goes back when a version without that name is installed, from the copy that update kept. Only the last three updates keep theirs, and the one that replaced this file is older than that: Parseh’s file of that name is deleted like any other the new version does not ship, and yours can come back only from a backup of your own (Backups).

3. The guide#

The guide's front page, served by Parseh before its first compile — The guide’s front page, served by Parseh before its first compile

What you see	What it means, and what to do
The front page says The guide has not been compiled yet.	The pages you are reading are Markdown until they are compiled into web pages. When Parseh serves the guide (the hub’s guide button), the same notice has a Compile the guide button: press it, and the new pages open when it is done. Opened from the disk, the guide cannot compile itself: run the installer, which compiles it every time (`./install.sh --guide` compiles only the guide).
The guide was compiled from older pages. Compile it again to see the newest.	A page of the guide changed since the last compile — after an update, say. Press Compile the guide again.
The last compile failed. What it said is below. or The compile ended with errors. Opening the pages it wrote…	The compile wrote every page it could and stopped on an error: each `warning:` or `error:` line below names the page and the line. That is a fault in a page of the guide, not in your content; the other pages work.
The search needs the compiled guide (site/search-index.js), and it is not there yet. and The list of pages appears once the guide is compiled.	The same thing: nothing has been compiled yet. See the first row.
`https://<account>.github.io/Parseh/` — a copy of your own — answers 404 or still shows older pages.	GitHub Pages publishes what is committed on `main`: switch it on once (Settings → Pages → Source: Deploy from a branch, branch main, folder / (root), Save) and give it a minute or two; a changed page appears after it is compiled and `html-guide/site/` committed and pushed. Compiling and publishing has the whole setup.
`build.py` says `error: … is not empty and … nothing was touched` and exits with status 2.	A compile replaces its folder whole, so `--out`, `--pages` and `--export` only go into a folder that is new, empty or one they made themselves before; each says it in its own words — `--out`: `does not hold an earlier compile of the guide (no build.json of its own)`; `--pages`: `was not laid out by --pages (no .parseh-guide-pages in it)`; `--export`: `does not hold an exported guide`. Name another folder.

4. Books#

What you see	What it means, and what to do
PDF behind the text — build it in the reader’s header.	You edited a chunk. The reader was rebuilt with your edit at once; the PDF is LaTeX, and is built only when you ask. Click the notice, or build PDF, and follow the build to its end.
…: the build failed — … after build PDF, build or rebuild on a card.	The message quotes the first error TeX gave (a line starting with `!`), or the line saying what could not run. no lualatex on this machine: the PDF needs TeX Live (or MiKTeX); building the reader alone means just that: the reader is built, the PDF needs TeX. `./install.sh --pdf` adds the TeX packages a PDF needs; on Windows, your TeX’s own package manager does.
The reader refuses to empty a box: … a glossed chunk needs its meaning …; emptying every box of the gloss at once (“delete gloss”) takes the whole gloss off	A box the language requires — the meaning, the transliteration where every chunk is romanised, a Japanese chunk’s kana — cannot be emptied on its own, since that would leave a finished gloss half finished. To start the chunk again, press delete gloss in the chunk sheet, which empties every box at once (and undo delete puts it back). See Deleting a gloss.
fill from the answer, in the reader’s gloss with an LLM sheet, lists a chunk under kept or dropped	kept: the answer tried to change something an answer may not, and that part was left as it is — the line says which: a gloss already there, the word line, the colour, a note, the free mark, or a plain chunk. It is not always a chunk left untouched: a chunk the answer filled is listed too when the answer also changed its word line or its colour — the gloss went in, the other change did not. To have an answer replace a gloss already there, tick re-gloss what is already glossed before you paste (or press delete gloss on the chunk first). dropped: the answer could not be written there, and the line says why — its text or its division differs from the book’s, it lies outside the stretch you picked (a folded paragraph among them), it would leave the chunk half glossed, or the chunk sheet’s own check refused a value. See Glossing a stretch with an LLM.
The reader refuses an edit: this text would stop paragraph N reproducing source/paras/…, which verify_book.py checks	A book’s chunks, joined, must still say what the text it was made from says. Adding vowels, harakat or punctuation is free; changing a word is not. If the change is meant — you are correcting the source itself — tick this paragraph need not reproduce `source/paras/` in the chunk sheet: the check then stands aside for that paragraph alone.
A feature this guide describes is missing from a reader, or the download sheet sizes all of it as far smaller than it is.	What a reader can do is written into the page when it is built, so a book built before a feature has not got it — and a reader built before the size fix says the size of the first recording only. Press rebuild the reader in its header (a few seconds, no LaTeX), or rebuild on its card in the library.
The library card says no audio yet for a book that has a narration.	The book came in from a bundle made with everything but the recording, which leaves the recording itself out. Put its `audio/` folder back in the book’s own folder, or add the narration again from the reader’s narration panel.
An upload on the Bring a book back panel says that is not a Parseh bundle: it carries no parseh-bundle.json. The download button makes one; a zip of a folder is not one.	Only a zip made by download (or by Backup every book) carries the manifest that lets Parseh check it. Unzip, edit, and zip the same folder again with its `parseh-bundle.json` at the top, or download the book afresh.
the bundle is damaged or locked and cannot be read	The download was cut short, or the zip has a password. Download it again.
An upload is refused with verify_book refused this book, the text no longer reproduces its own source paragraphs, or … does not parse	The bundle was checked before anything was written, and one of its chapters breaks a rule: the message quotes the place. Fix it in the files and upload again; nothing on the shelf was touched.
An upload answers that the book is already here, with a choice to replace it.	A bundle never overwrites a book by surprise. Replace it, or cancel and keep the one on the shelf.
cut the audio… is greyed out, with a reason under it.	The reason is the answer: this book has no narration file to cut the audio from; this subparagraph has no times yet: time it (narration, or edit times) and the audio can be cut; none of this book’s recordings covers … yet; … was timed in the recording …, whose file is not on this machine; cutting the audio needs this page opened from the toolbox’s server.
The cut editor draws no waveform, and the clip comes out as a `.wav`.	The machine serving Parseh has no `ffmpeg`, so the clip is recorded in the browser instead of cut by the server. It still works. Install `ffmpeg` (the installer’s report says whether it is there) for the waveform and MP3 clips.
by the sound, beside estimate the rest in a recording’s by ear, is greyed out.	Its tooltip says why. this reader was built before estimating by the sound existed…: press rebuild the reader in the header (a few seconds, no LaTeX), then reload the page. there is no picture of the sound here to estimate by — it is drawn by ffmpeg, on the computer Parseh runs on: the machine serving Parseh has no `ffmpeg` (the installer’s report says whether it is there). by the text works either way (Fixing the timings).
estimate the rest by the sound, in a book or a video, says estimating by the sound needs numpy, which the Python running Parseh does not have…	The environment lacks one package. Run the installer again (it adds what an environment lacks), then start Parseh again. Until then by the text works.
estimate the rest by the sound says another estimate by the sound is running: try again when it has finished	One runs on the computer at a time, and one is running — perhaps one you gave up with stop estimating, which the computer still finishes. Wait a little and press E again (its Working… entry is gone when it is done), or go by the text.
… or this stretch is too long to estimate by the sound at once: estimate from a line nearer the end, or by the text	The computer ran short of memory for a stretch that long. Estimate a shorter stretch — from a line further on — or by the text.
The server’s console says `!! … declares a narration at …, which is missing`	`book.json` names a recording that is not in the book’s `audio/`. Put the file back, or open the reader’s narration panel, which says the same and lets you add it again.
A book taken off the shelf with ✕ is gone from the library.	Nothing was deleted: the whole folder was moved to `books/.trash/`, named after the book and the minute. Move it back into `books/<language>/` and write the library page again: rebuild on any book’s card does, or `python3 lib/make_index.py`.

5. A book made by an agent#

What you see	What it means, and what to do
open the folder is not there on another device, and the panel says Opening the folder shows it on the screen of the computer Parseh runs on…	Opening a folder is the computer’s own act: it shows it in the file manager on that computer’s screen. Everything else — making the folder, giving the agent text, this is all the text, finish, reopening — is done from any device let in; here, copy the folder’s path.
finish… lists part 3 has not been taken by the agent yet or the agent has not said it is done	Finish says what it would not wait for: the agent has not taken every part you gave it, or has not said it is done (its stage is not all batches in or waiting for the next part). Tell the agent to carry on, or press finish anyway if it forgot to record.
The agent does not take a part you added.	It reads `making.json` and `ASKS.md` before each batch, so tell it to carry on (a part is noted in both). A folder made before text could come in parts has instructions that do not mention them: tell the agent to read `parts` in `making.json`, or make the folder again.
choose the original first	The file picker is empty. Choose the PDF, the epub or the text file the book is made from.
a book is already at books/english/mini-en/ — choose another slug	A book with that name is on the shelf, or an earlier folder was made for it. Give the new one another slug, or take the other off the shelf with ✕ on its card (it goes to the trash, not away).
the original has to be a PDF with a text layer, an epub or a plain text file (.pdf, .epub, .txt): ‘letter.docx’ is none of them, or letter.pdf does not look like a PDF: choose the right file	The original is one of three kinds, and its first bytes have to say what its name says. The agent reads text, so a PDF that is only pictures of pages — a scan — has nothing to read: save the text as a plain text file, or choose a PDF that has a text layer.
This book is being made by an agent … editing is off until the making is finished, or the pencil says editing is off while an agent makes this book	The agent writes the book’s `.tex` from its own files (`annot/`), so anything you changed by hand would be erased by its next batch. Ask instead: the making panel has a box, what to change from now on, and the chunk sheet has ask about this chunk. When the making is finished, the reader edits again.
The reader says More is written than this page shows, or the panel says The agent has written more since this page was built	The page is a reader built earlier, and the agent has written since. Press look at it now in the making panel: it builds the reader again from what is there.
The panel says Parseh was updated during the making: it began under … and this is …	Parseh was updated while the agent worked, and the tools it calls — the checkers, the assembler — may have changed under it. Nothing is stopped or changed. Read what the versions in between changed (What changed, version by version), and write an ask if the agent should know.
The panel says making.json cannot be read just now	The agent is writing the file at that moment, or wrote it half. The book stays locked and being made; the panel is drawn again in a few seconds. If it stays, tell the agent to write `making.json` again.
The panel says The agent has not written anything yet for a long time	Nothing has opened the folder. Open it in the agent you use (open the folder, or the path above it) and tell it to read `AGENTS.md` and begin. Parseh does not start an agent.
the PDF of these chapters is not available on this computer, which has no shell to make it. The reader is available: look at it now	The PDF of the chapters so far is made by a shell script (`build.sh --draft`), and this computer — Windows above all — has no shell. look at it now builds the reader, which shows everything written; the whole book, PDF included, is built by finish and by build PDF, which have a Python build for a computer with no shell.
finish… says N paragraphs do not reproduce their source, with chapter 1 paragraph 3 does not reproduce its source (it differs at character 41) under it	The text of the chunks, joined, has to give back the original paragraph. Nothing is finished. Tell the agent which paragraph (an ask, then let it repair the batch), or, if the source itself is what is wrong, do it in the reader once the book is finished, with this paragraph need not reproduce `source/paras/` (What a book is made of).
finish… says there is no chapter in the book yet, so there is nothing to finish	The agent has not written a chapter. Finish is for a book whose batches are in. To give the book up, take it off the shelf with ✕ on its card.
finish… says the build failed: … with TeX’s first error after it	The check passed and the whole build did not. The message is the first line TeX gave (one that starts with `!`): it names a file and a line. Mend it — or ask the agent to — and finish again. The book is still being made.
finish… says the reader was built; the PDF was left, because TeX is not installed on this computer	The book is finished, and it has its reader. The PDF needs TeX Live (or MiKTeX): install it, and build PDF in the reader makes the PDF.
A card still says being made and nobody is making it	The agent stopped, and nobody pressed finish. Finish it if the batches are in, or take it off the shelf with ✕. A folder Parseh made stays being made until one of the two is done: Parseh never guesses that an agent has given up.

6. Videos#

What you see	What it means, and what to do
cut the audio… is greyed out on a YouTube video: only Chrome and Edge, on a computer, can record the sound of a tab (or this browser cannot read the sound of a shared tab)	A YouTube video’s sound reaches no script, so the clip is recorded from the tab as it plays — which only Chrome and Edge on a computer can do, and not a very old version of them (the second message). Open the video in an up-to-date one. On a phone the button is always grey. A film on this machine has none of this: its clip is cut from the file, in any browser.
recording this tab needs the page opened at the toolbox’s https address, in Chrome or Edge	Open the page at one of the `https://` addresses Parseh printed.
the tab was shared without its sound: press “record” and, when Chrome asks, leave “Also allow tab audio” turned on	In Chrome’s question, the switch that shares the tab’s sound was off. Press record again and leave it on.
no sound was captured — is the video muted? Turn its sound on and record again	Nothing but near-silence was recorded between the two ends. Check the video can be heard there, and record again.
by the sound, beside estimate the rest in the timings, is greyed out on a YouTube video: draw the sound first (“● draw the sound”, above the picture): there is no picture of it here yet to estimate by	A YouTube video has no picture of its sound until one is recorded from the tab. Press ● draw the sound in the same sheet; once it is drawn, by the sound can be pressed. Where the browser cannot draw it, the tooltip says so and why (The timings). A film on this machine needs only `ffmpeg`, as a book does.
Saving a phrase in the player is refused: … chunks do not reproduce the text, with the text and the chunks under it.	A caption’s phrases, joined, must give back the caption word for word, and the caption must say what its line of `transcript.txt` says: that check is what stops a model, or a careless edit, quietly rewriting the video. So adding vowel marks is free, and changing a word is refused. If the change was a slip, undo it. If the transcript itself is wrong — YouTube misheard a name, dropped a particle — tick this phrase need not reproduce `transcript.txt` at the foot of the phrase’s ✎ edit form (the refusal scrolls it into view) and save again with the correction: the edit is written, the caption’s text follows its phrases, and that caption is no longer checked against the transcript. See When YouTube heard wrong.
Unticking this phrase need not reproduce `transcript.txt` is refused: segment N: text differs from transcript.txt	The caption still says something its line of the transcript does not. Put its words back as the transcript has them in the same save as the untick, or leave the box ticked.
The video checker warns `N captions not checked against transcript.txt: a chunk of each is marked as departing from it`	Not a fault: it counts the captions freed with the box above, so that none is forgotten.
A caption cannot be cut, joined or deleted in the player.	On purpose. The captions are what the checker holds a video’s two files to, caption for caption, so once the video is added only a caption’s start moves (the timings, in the player’s bar). Its phrases can still be cut and joined — ✂ cut in two, join next, join previous in the ✎ form — and its words corrected as above. To re-cut the captions themselves, mend the transcript before adding (Edit the transcript… on the add page), or download the video with ⤓ in the player’s bar, edit the files, and bring the zip back through Bring a video back on the video index.
The video checker says `segment N (start S) chunk C: missing 'tr'` (or `'en'`, `'kana'`), and `note: N of M chunks have no gloss yet`	The first is a phrase half glossed — something written in it, something its language requires still empty: an error to the checker, though the ✎ form saves it on the way. Fill the missing box in the player. The note only counts the phrases nobody has glossed yet, which are legal in every video (A video still being glossed). A `"draft": true` in an older `video.json` changes none of this: nothing reads it.
Emptying a box in the ✎ form is refused: … a glossed phrase needs its meaning …; empty every box of the gloss (“delete gloss”) to take the whole gloss off	A box the language requires cannot be emptied on its own, since that would leave a finished gloss half finished. To start the phrase again, press delete gloss, which empties every box at once — and undo delete if that was not what you meant. See Deleting a gloss.
fill from the answer, in the player’s gloss with an LLM panel, lists a phrase under kept or dropped	kept: the answer tried to change something an answer may not, and that part was left as it is — the line says which: a gloss already there, the word line, the colour, a note, the free mark, or a plain phrase. It is not always a phrase left untouched: a phrase the answer filled is listed too when the answer also changed its word line, its colour or its note — the gloss went in, the other change did not (a correction of the caption belongs in the meaning or the vocabulary line). To have an answer replace a gloss already there, tick re-gloss what is already glossed before you paste (or press delete gloss on the phrase first). dropped: the answer could not be written there, and the line says why — its text or its division differs from the player’s, its start is another caption’s, or it would leave the phrase half glossed. See Glossing captions with an LLM.
A colour or a gloss you wrote by hand into a video’s annotations file has gone.	This cannot happen to a video added on or after 23 September 2026: the batches are folded in once, while the video is being added, and then dropped (parts/ is retired). On an older video that still carries `parts/`, `merge_parts.py` run by hand rebuilt the file from batches that never had your edit — it now refuses instead, naming what it would lose. If it is already gone, the copy in `.trash/` or a download made with ⤓ is where to find it.
A video taken off the index with ✕ is gone.	Nothing was deleted: it was moved to the `.trash/` folder under `youtube/videos/`, named after the video and the minute. Move it back into its language’s folder beside `.trash/`.

7. Cards, clips and Anki#

What you see	What it means, and what to do
An edit you made inside Anki is gone.	A rebuilt deck was imported without syncing first: the build writes every card as the store has it. Redo the edit in Anki, then always sync first (Anki on the hub, steps 1–4) before you rebuild and import.
The sync wizard says A package built by Parseh … not an export from Anki. It says nothing about what Anki holds, so there is nothing to sync	You dropped the `.apkg` the build button made, not an export from Anki. It can still be brought in (Bring this deck in, Add the N new cards), but to sync, export the deck from Anki: File → Export…, Anki Deck Package (.apkg).
The sync says `that file is not an Anki deck export -- export with File > Export > Anki Deck Package (.apkg)`	The file is not a deck package. Export again from Anki as the message says.
this export is in Anki’s modern compressed format, which needs the zstandard package	The environment lacks one package. Run the installer again (it adds what an environment lacks), or export from Anki again with Support older Anki versions ticked.
Building a deck refuses: the note type … in your Anki collection has fields …, but this exporter writes …	The note type gained or lost a field inside Anki. Building now would fork a second note type and orphan your review history, so it refuses. The exporter has to be taught the new shape first: the card store’s own `README.md`, in `youtube/anki/`, says how.
Here, not in Anki yet in the sync’s preview.	Harmless: cards made on these pages since your last import. Anki has never seen them; step 5’s import gives them to it.
Images and recordings not in this export	Export from Anki again with Include media ticked, or those pictures and recordings go missing from the cards.
A card vanished from a deck here.	Look in the deck’s `deleted/` folder, in the card store (`youtube/anki/`, then the language and the deck): a sync that removed cards you deleted inside Anki moves their files there. Move a file back into `cards/` beside it to undo.
A card pasted into a document as markdown shows no picture or plays nothing.	Its files wait in the clip tray, and come into a document when it is saved. Save the document. If the clip was deleted from the tray (The clip tray on the hub) in the meantime, cut it again.

More in The round trip with Anki and The clip tray.

8. Reading help: dictionaries, sentences, the translation model#

What you see	What it means, and what to do
No dictionary switch in a reader’s header.	Nothing is installed for that language yet. Open Get a dictionary for this language → in the ⚙ panel’s Looking a word up (or Settings → Reading help) and press Get it on the language’s dictionary.
The panel gives dictionary entries but never a sentence.	No parallel corpus is installed for that pair of languages, or nothing in it shares a rare word with your chunk. Coverage is very uneven; a thin pair finds a sentence less often. Get one on the language’s card of the reading-help page, the row Sentences people translated; Japanese and Chinese need their dictionary first.
The panel never shows a machine’s reading.	There is no button for it: it is drawn when the panel opens or not at all. Either no model is installed for the pair — models go to and from English only — or the dictionary switch is off, which is the switch the reading rides on.
the model could not be run	The model’s files are incomplete, or the page is served by something other than Parseh (which names `.wasm` files as the browser requires). On the reading-help page, Remove… the model and Get it again.
The sources sidebar offers a verb as `\dw`, not `\vb`.	Its language’s recipe did not take that hit for a verb it can fill in — a Chinese verb that neither splits nor takes a complement, a word the dictionary lists only as another verb’s form — or the dictionary predates the recipes. Press Rebuild on the language’s dictionary, on the reading-help page.
A character chosen after Decompose Kanji (or Decompose Hanzi) says Install character components.	No component pack is installed. Open dictionary & component setup leads to the language’s Character components row on the reading-help page: Get it on KanjiVG (Japanese), Make Me a Hanzi (Chinese) or CJKVI-IDS (both, on the card every language shares).
The tree says No component decomposition is available for this character.	The installed packs do not cover it. The fallback pack, CJKVI-IDS, covers the most; Manage component packs and fallback coverage leads there.
The tree’s foot says Install the Japanese dictionary to show component meanings. (or Chinese)	The tree works without a dictionary; the meanings of the components come from it. Open setup and Get it on the language’s dictionary.

9. Studio#

The name-conflict dialog, with a name that is taken too said under its field

What you see	What it means, and what to do
A dialog: This name is already in use	A link names a document by its title, so no two documents of one library may share one. The dialog shows each clash with two fields — Rename the document already in the library (with open it ↗ to look at it first) and Rename the document being saved (or added). Change either or both and press Use these names; every link to a renamed document follows it. Cancel writes nothing, and the editor keeps your text. It opens on Save, Save & view, Ctrl+S, a new document’s first save, Paste LLM answer, Upload .md / .zip and Load from backup.
Under a field: “X” is already the name of another document — choose another name	The name you typed is taken too. Type another; the dialog stays open.
A document needs a name — give it one or A name cannot hold “\|”: a link to it in a table would split the cell — choose another name (dialog: This name cannot be used)	A name may not be empty, and may not hold a `
A link drawn in red and dashed, whose tooltip says No document named “X” in this library — create or upload one with this name and this link will work again	The document it names was deleted, renamed outside Parseh, or not written yet. Links are never removed: make or upload a document with that title, or rename one to it, and the link works again with nothing rewritten.
A document’s title now ends in a number — A note 2, A note 3 — and you did not rename it.	Names became unique with links by name, and a library written before could hold two documents of one name — a book’s notes were all called A note. The first time the server met each library after the update (the studio’s when it started, a book’s or a video’s notes when they were first opened), it kept the name for the oldest document and put 2, 3… after the name of each later one, in its header; the server’s console said so. Links that named it still find the oldest. Rename it as you like — change its `title:` line in the editor and save — and see Links written before names.
An exercise drawn as a box: Exercise needs attention. with a list under it.	The list names each problem in the exercise’s markdown — a field that is not known, a missing answer, a picture that is not under `images/`. Fix them in the editor; ✎ Edit on the exercise in the preview reopens its form. Such an exercise has no + Deck button until it is fixed.
The build badge says build failed — see message.	Point at the badge to read what LaTeX said. xelatex not found — install TeX Live means just that. fonts not provisioned yet happens only in the first moments after the server starts: wait a moment and build again.
Large print needs the TeX package extsizes (extarticle.cls), which this TeX installation does not have.	The larger print sizes need one TeX package. On Linux and macOS, `./install.sh --pdf` installs it. On Windows, where that script does not run, add the package `extsizes` with your TeX’s own package manager — the Packages page of MiKTeX Console, or the TeX Live Manager. Meanwhile build at Normal · 11 pt.
source changed — rebuild beside a PDF.	The markdown changed after the PDF was built — perhaps only because another document was renamed and a link in this one followed it. Build PDF again.
built with other options — rebuild	PDF options ▾ now says another print size or colour than the PDF was built with. Build PDF again to use them.
N RTL FAIL (or N text FAIL), not verified, N overfull	The PDF was checked against the markdown: the first lists the strings of the target language that did not come out right in the PDF; not verified means PyMuPDF, which does the check, is not installed (run the installer); overfull counts lines that stick out into the margin.
An upload opens a dialog saying Its header has no …	A studio document’s header needs `title`, `subtitle`, `note`, `lang` and `target`. The dialog asks for exactly the missing ones and adds them; the rest of the file stays as it is.
Load from backup says that zip holds no library (or that is not a zip file)	The file is not a library backup: that is what Backup writes. A single document’s zip goes in through Upload .md / .zip instead.
A picture or a recording is refused on upload: only PNG, JPEG, SVG and PDF figures are supported, image larger than 15 MB, only … recordings are supported	The studio takes those picture formats, and the recording formats the message lists (MP3, M4A, AAC, Ogg, Opus, WAV, FLAC, WebM). Convert the file and upload again.
A starter’s picture or chime you deleted shows as missing, and is not copied back.	By design: once deleted, a starter’s file stays deleted for that document. Delete the line that names it, or upload a file of your own under that name.
A page made with Download ▾ → HTML page, for a website, opened from the disk, shows Video player configuration error where its video should be.	YouTube plays nothing for a player that does not say which site it is on, and a file opened from the disk is on no site. The line under it says so too, and watch it on YouTube opens the video there. Put the page on a website and the video plays in the page; everything else on the page works from the disk as well. See A page for a website.
The editor writes right to left and you did not ask for it.	The document’s `lang:` is a right-to-left language, so the editor opened in that direction. Press ⇥ LTR editor to switch; the choice is remembered for this document.

More in Names and links, Printing to PDF and A page for a website.

10. Exercises#

What you see	What it means, and what to do
this exercise was already answered — showing what comes next	You answered the same exercise in another tab (or a request was sent twice). It is not scheduled a second time; the page simply moves on.
the document has changed since the page was loaded: reload it, then copy	The document was saved elsewhere after this page loaded, so + Deck could pick the wrong exercise. Reload the page and copy again.
this exercise is in X, but the deck “Y” is for Z	A deck has one language. Choose (or make) a deck of the exercise’s language. Move to deck… says the destination deck must have the same language for the same reason.
Import deck… says this is not a deck export: parseh-exercise-deck.json is missing or the file is not a zip	Import deck… takes the zip a deck’s Export makes. A zip of every deck, made by Backup on the Exercises page, goes back through Load from backup beside it.
… is not a recording: only … recordings can be imported, the zip is larger than 512 MB uncompressed	The import checks every file before it writes anything. Remove the offending file from the zip, or export the deck again.
An exercise has no + Deck button.	It shows errors (fix them first), or you are in the editor’s preview, which has none: open the document’s reading view.
A flashcard does not play its recording by itself.	Many browsers let a page make no sound until you have clicked on it. Click the card’s 🔊 once; from then on the recordings play.
Study now shows next exercise in … or no exercises scheduled	Nothing is due. That is the scheduler working: come back then, or use Cram exercises on the deck’s page, which never changes the schedule.

More in Studying and Export, import and backup.

11. Working…, and the Browser | Mobile switch#

What you see	What it means, and what to do
The Working: pill in the corner stays on Preparing the download…	The server is packing the zip — a book with its recordings, a whole shelf — before it can send it. The pill goes once the file has been sent; the file itself arrives in your browser’s downloads, as always.
The pill says Working: but nothing seems to happen on the page.	The work may have been started on another page or in another tab: the pill shows everything the server is doing. Click it for the list; open the page goes to the page each one was started from.
Something was refused, and the pill said nothing.	On purpose: the pill says only what is running. A refusal is said by the page that asked, as a message or a dialog.
In Mobile mode a door opens the ordinary page.	The hub, the books and the exercise decks have their mobile versions; the videos and the studio’s notes open their browser pages until their own mobile pages are written.
In Mobile mode a book’s reader has no book info, build, narration or pencil, and an Alt-click makes no card.	On purpose: the mobile reader is for reading. Press Browser in the ⚙ sheet’s Interface group and they are all back, on the same page.
The mobile shelf says a book is not built yet, and its card opens nothing.	Its reader has never been built. Build it from its card in the library, in the Browser interface.
The hub has no Anki, clip tray, reading help or stop button any more.	The hub is in Mobile mode, which leaves out everything that edits or administers. Press Browser in its top bar.
Studying in Mobile mode, there is no bar at the top and no Deck button.	On purpose: the exercise has the screen. ‹ beside the deck’s name goes back to the deck, and ⚙ beside it opens the page’s panel.
In the ⚙ panel, the Page zoom steps above 110 % are greyed and say this screen is too narrow for more.	The page would be narrower than 320 pixels at that size. A phone held upright (about 390 pixels) stops at 110 %; turn it sideways, or use a wider window (Page zoom).
Page zoom is greyed out whole and says this browser cannot zoom the page.	That browser’s own zoom is not one Parseh can use; the page is left exactly as it was. Use the browser’s own (Ctrl +, Ctrl −).
In a Persian or Arabic book there is no Diacritics row.	The row is only there for a language that has vowel marks to put away — in a video, only when its lines hold some (Levels and their names).
A flagged chunk is listed with this chunk is no longer where it was, and ▶ will not go.	The book was edited since, and the chunk is not there any more. It stays so that nothing is lost: take it off with ✕, or copy it (Review later).
I flagged chunks on the phone and the computer does not list them.	A phone keeps its flags and sends them to the computer when it can reach it: the list says … will be sent when the computer can be reached until it has. Open any Parseh page on the phone with the computer on.
As an app says the phone does not trust Parseh’s certificate yet, or the phone’s menu only offers a shortcut that opens with an address bar.	The phone has to be told once to trust Parseh’s certificate: step 1 on the As an app page — download it, install it in the phone’s settings (on an iPhone, turn on full trust for it as well), then open Parseh again.
The app says Parseh cannot be reached.	The phone cannot reach your computer: it is asleep or off, Parseh is not started on it, or the phone is away from its network (and from Tailscale, if that is how it reaches it). Start Parseh on the computer, then Try again — it says Trying… while it waits, and Still cannot reach it. Tried again at … if the computer still does not answer. Go back returns to the page you came from.
A link to a page I never kept waits, then says Parseh cannot be reached, although the computer is on.	The computer is waited for as long as it answers the Parseh page you are on — up to half a minute, the page you are on staying as it is. The message comes early only when nobody could say that it answers: after about four seconds if the computer answered nothing (asleep, suspended, off the network) or if the Parseh pages that are open could not say — a tab behind another, and the cannot be reached page itself, where Try again is pressed; after two and a half if Parseh had already said the computer was away, or if no Parseh page was open at all — a link typed in a new tab or opened from another app. Go back, and follow the link again from a Parseh page that is in sight: it waits.
A page says offline although the computer is on.	Press ↻ beside the chip, or pull the page down from the top: a refreshed page asks again, and opens online if the computer answers. A page that went offline stays so while it is open, so that a book does not change under you — and Parseh says offline only when the phone has no network, when the computer refuses (Parseh not running on it), or after 45 seconds without any answer.
checking… stands where the offline chip goes.	The computer is slow to answer, not gone: the page goes on as usual, and the words go as soon as it answers. Only if nothing answers for 45 seconds does it say offline.
Keeping a book says Chrome will not download for Parseh in the background — turn on Automatic downloads for this site in Chrome’s Site settings.	Chrome refused to bring the book by the phone’s own download. Nothing was fetched. Allow Automatic downloads for Parseh in Chrome’s Site settings, then keep it again.
A keep ends with …: stopped from the notification — what had come is on this phone, …: 2 of its files are no longer on the computer or …: no room on this phone for 2 of its files	The book’s download stopped part-way, for the reason it names: Cancel on its notification (or Stop on Kept on this phone, which says stopped on this phone), files the computer no longer has, or a phone that filled up. What had come stays, and the book is kept with it; Change what is kept shows what is missing, unticked. Tick it and Save to fetch the rest — after giving some room back on Kept on this phone, if it was the room; a file the computer no longer has cannot be fetched again, from here or anywhere. After a first keep, — nothing of it had come yet means the book was not kept at all: press Keep on this phone again.
Keeping a book says … was already coming onto this phone — change what is kept once it is in	Another tab had started keeping the same book after this one’s list was opened. That keep carries on, and what it brings is kept; what was ticked here was not fetched, and nothing was given back. Once it is in, open Change what is kept again.
The button says Waiting for … — you can leave this page	Another book is coming onto the phone, and this one waits behind it and starts by itself once that one is in. Nothing to do. With two already coming, one more may be refused instead: Two books are already coming onto this phone — keep this one once one of them is in.
The button says Keeping… 43% — waiting for the computer	The computer stopped answering in the middle of a keep: asleep, off, or out of the phone’s reach. The download waits, and carries on by itself when the computer answers again; a book waiting behind it waits too. Wake the computer, or start Parseh on it.
After an update, every browser warns about Parseh’s certificate once more.	An update from Settings keeps the certificate, `.tls/`, as it was, so this is the one-time change of a Parseh from before the certificate had an authority of its own, which a phone can be told to trust: the first start after it made the new one. Accept it once, as before — or, on a phone, trust it for good from As an app. A fresh install that did not get the old folder’s `.tls/` makes a certificate of its own too (Moving into a fresh install, once).

More in The Working… indicator and Browser and Mobile.

12. Skills for your chatbot#

What you see	What it means, and what to do
The chatbot answers a request with I do not have the parseh-gloss skill.	The skill is not installed in that chat. Install it (Skills for your chatbot), or use copy the prompt, which needs no skill.
The chatbot says the request is for a newer skill than the one installed.	You updated Parseh after downloading the skill. Download it again from Settings → Skills for your chatbot and install it over the old one.
The row says your skill may be older than this Parseh — download it again.	This device remembers the last skill it downloaded, and Parseh would make another one now. Download it again.
copy the request for the skill is off and says the prompt takes the place of Parseh’s.	A prompt of yours that is in place of Parseh’s cannot travel in a request: the skill carries Parseh’s instructions. Use copy the prompt, or choose Parseh’s own or an added prompt.
A skill uploaded to claude.ai is not used.	Check the skill’s switch in Customize → Skills, and that code execution is on in Settings → Capabilities; on a Team or Enterprise plan an Owner decides (Skills for your chatbot).